County-level Vulnerability Assessment for Rapid Dissemination of HIV or HCV Infection Among

County-level Vulnerability Assessment for Rapid Dissemination of HIV or HCV Infection Among Persons who Inject Drugs, United States Supplemental Appendix Contents Supplemental Methods... 2 Regression Modeling Analyses... 2 Modeling Procedure... 3 Continuous Indicators Linearity Assessment... 3 Collinearity Assessment of Indicators... 4 Standardized Regression Coefficients... 5 Composite Index (Vulnerability) Score and Rank... 5 Supplemental Results... 6 Model Fit Results... 6 Composite Index (Vulnerability) Score and Rank... 6 Counties Identified as Vulnerable... 7 References... 7 Tables... 7 Table S1. Counties identified in the top 5% of vulnerability ranks by state and rank... 8 Table S2. States with at least one county identified in the top 5% of highest vulnerability ranks by number of vulnerable counties and percentage of all state counties identified as vulnerable.... 9 Figures... 9 Figure S1. County-level indicators investigated for association with acute HCV infection.... 9 Figure S2. Acute HCV infection rate by county. Reported rate of acute HCV infection by county, NNDSS 2012-2013 and model-estimated rate of acute HCV infection by county... 18 Figure S3. A. Histogram showing vulnerability scores by number of counties in each score group (eg, 0 to 0.25), B. Sigmoid curve showing vulnerability scores by county rank and identifying the scores bordering the top 5% cut-off, and C. Caterpillar curve of vulnerability scores by mean rank 90% confidence intervals bordering the top 5% rank cut-off (red dash line).... 19 1

Supplemental Methods We identified indicators associated with acute hepatitis C virus (HCV) infection to develop a composite index score (vulnerability score) ranking each county s vulnerability to rapid dissemination of IDU-associated HIV if introduced, and new or continuing high numbers of HCV infections among persons who inject drugs (PWID). We chose acute HCV infection as the outcome that best serves our purpose because it is collected at the county-level for almost all states. Regression Modeling Analyses We modeled the number of acute HCV infections by county using a multilevel Poisson model with the county population set as the offset. 1 Our data have a multilevel structure with the i th year (2012, 2013) nested in the j th county and j th county nested in the k th state. The Poisson distribution is defined as: where y = 0, 1, 2,..., and λ is the expected rate. The Poisson model uses the loge function that relates the expected value of the response variable to the linear predictor. Hence, the expected rate, λ, is modeled using the link function loge as: where X is the i th indicator and β is the associated model parameter, and β0 the intercept (i.e., overall mean). The offset, loge(population), is the county population. We have multilevel data and we model the levels (i.e., state and county nested within a state) as random effects to account for spatial heterogeneity (i.e., overdispersion). We modified our model so the loge link function relates the conditional mean (i.e., conditional on the random effects) of the response variable (i.e., acute HCV rate) to the linear indicator of the fixed and random effects. Our model including random effects for the county and state is given by: 2

Where the random effects, b0 and b1, are assumed to be N(0, σ 2 jk) and N(0, σ 2 k), respectively. We used SAS GLIMMIX 2 and the residual subject-specific pseudo-likelihood (RSPL) model estimation method. Modeling Procedure We fit a univariable Poisson random-effects models for each of the 15 considered indicators. Figure S1 depicts county-level data by class for each of the 15 considered indicators. Per capita income and population density were modeled on log10 scale. Aside from urgent care and highway exit, which were coded as yes/no, the other indicators were treated as continuous variables. Our goal was to develop a parsimonious model that is significantly associated with acute HCV infection rate. We entered all 15 indicators in the multivariable model and removed the indicators with the highest p-value. We removed and added indicators in a backwards stepwise procedure until all remaining indicators had a p- value<0.05. Continuous Indicators Linearity Assessment We assessed linearity for the 13 continuous indicators. Our assumption was that these indicators were linear on the loge(acute HCV rate) scale. To assess the assumption of linearity of the rate on the log-scale we used the following procedure. 1. Calculate the quintiles for the indicator 2. Calculate the acute HCV rate by quintile 3. Plot the loge(acute HCV rate) versus the quintile for the indicator 3

4. Estimate the slope and intercept of the loge(acute HCV rate) versus quintile 5. Visually assess the assumption of linearity of the indicator Collinearity Assessment of Indicators If an indicator is nearly a linear combination of other indicators in the model, the affected estimates may be unstable and have high standard errors. This situation is usually referred to as collinearity or multicollinearity. We used a generalized linear model (GLM) with counts as the outcome, which required a different procedure to assess collinearity than for a linear regression model. To assess collinearity we relied on three calculated statistics: eigenvalue, condition index, and principal component proportion of variation. An eigenvalue is a computed value that characterizes the essential properties and numerical relationships within a matrix. Eigenvalues that are close to zero may be indicative of a matrix that is close to singular, which indicates collinearity. Eigenvalues <0.01 are usually thought to be close to zero. The condition index is defined as the square root of the ratio of the largest eigenvalue to each individual eigenvalue. The largest condition index (i.e., the square root of the ratio of the largest to the smallest eigenvalue) is the condition number of the scaled X matrix which, as noted by Belsley et al (1990), suggest that when this number is nearing 10, weak dependencies might start to affect the regression estimates. 3 When this number is larger than 100, the estimates likely include significant numerical error. To calculate the eigenvalues, condition indices, and proportions of variation we used a two-step process in SAS. 2 First, we fit the Poisson model using PROC GENMOD and output the Hessian weights. Secondly, we fit a linear model, using PROC REG, with the Hessian weights defined in the weight statement to obtain the collinearity diagnostics. 4

Standardized Regression Coefficients Standardized regression coefficients for our final multivariable model were calculated to determine the relative importance of each indicator. We calculated the standardized regression coefficients using: Where βp is the estimated regression coefficient from the final multivariable model, Std is the standard deviation of the X th indicator and pseudo outcome y. The pseudo outcome is the outcome estimated on the loge scale from the estimated regression model. Composite Index (Vulnerability) Score and Rank Our primary goal was to develop a composite index score for ranking the county vulnerability to rapid dissemination of IDU-associated HIV if introduced, and new or continuing high numbers of acute HCV infection among PWID. We developed a vulnerability score using data from the indicators identified in the final multivariable model and the following method to rank counties from lowest to highest vulnerability. We used regression coefficients and observed values to compute the index score for each county. The score for the j th county was calculated using the regression coefficients (β) and indicators (X) as given by: The intercept, β0, is not used because it is a constant and has no impact on the ranking of counties based on the scores. Once the vulnerability score was calculated for each county, including those not used in 5

fitting the model, they were ranked from 1-3143 with higher scores interpreted as being more vulnerable. Ranks using regression coefficients include uncertainty. To account for uncertainty in the ranks we used simulation to estimate the 90% confidence interval (CI) for each county s rank. We drew 10,000 samples from a normal distribution for each regression coefficient using their estimate and standard error of the estimate. For each of the 10,000 samples we calculated the county's vulnerability score and rank and then obtained a CI for each county's rank. The threshold for classifying the most vulnerable counties was set at the 95th percentile (top 5%). The 95 th percentile threshold of the ranks was calculated using all 3,143 counties as 0.95 * 3,143 = 2985.85. We used the upper bound of the 90% CI to determine if a county's rank was within the 95 th percentile. Once we determined the counties that were within the threshold we ranked them using the inverse of their mean estimated rank (1=highest vulnerability). Supplemental Results Model Fit Results The final multivariable model with the 6 indicators closely aligned the reported HCV rates with the model-estimated HCV rates. To illustrate the model fit we mapped the reported and model-estimated rates of acute HCV infection per 10,000 population (Figure S2). Fewer than 15% (469 of 3,143) of counties varied by more than 1 class when comparing the reported and model-estimated rate of acute HCV infection. The average absolute difference in the model-estimated rates was 0.011 per 10,000 population lower than the actual rates. Composite Index (Vulnerability) Score and Rank Figure S3a shows a histogram of the vulnerability scores by number of counties in each score group (e.g., 0 to 0.25). Figure S3b shows a sigmoid curve of the vulnerability scores by county rank. 6

Using the mean average rank, 157 counties were ranked above the inclusion threshold. The red circle identifies the counties that border the top 5% cut-off. Figure S3c shows a caterpillar curve of the 90% confidence intervals (CIs) bordering the top 5% cut-off. An additional 63 counties were identified above the threshold based on their 90% CI for a total of 220 vulnerable counties. Counties Identified as Vulnerable Table S1 lists the counties identified within the top 5% threshold of vulnerability ranks by state and rank. Table S2 summarizes information on the 220 counties by state; including information on the number of counties identified. Seven states had 10 or more vulnerable counties: Indiana, Kentucky, Michigan, Missouri, Ohio, Tennessee, and West Virginia. References 1. Gelman, A and Hill, J. Data Analysis Using Regression and Multilevel/Hierarchical Models. (2007). Cambridge University Press. 625 p. 2. SAS Institute, Inc. SAS : Version 9.3 for Windows. Cary (NC): SAS Institute, Inc.; 2012. 3. Belsley DA, Kuh E, Welsch RE. Regression Diagnostics: Identifying Influential Data and Sources of Collinearity. (1990). John Wiley & Sons Inc. Tables 7

Table S1. Counties identified in the top 5% of vulnerability ranks by state and rank FIPS County Rank FIPS County Rank FIPS County Rank FIPS County Rank Alabama Kentucky (cont.) Missouri (cont.) Tennessee (cont.) 01127 Walker 37 21133 Letcher 50 29153 Ozark 185 47063 Hamblen 138 01093 Marion 100 21115 Johnson 53 29229 Wright 194 47007 Bledsoe 139 01133 Winston 109 21207 Russell 54 Montana 47159 Smith 140 01059 Franklin 206 21063 Elliott 56 30061 Mineral 161 47109 McNairy 141 Arizona 21125 Laurel 65 30103 Treasure 211 47139 Polk 142 04015 Mohave 208 21041 Carroll 67 Nevada 47089 Jefferson 149 Arkansas 21217 Taylor 75 32029 Storey 52 47163 Sullivan 151 05135 Sharp 157 21081 Grant 77 32009 Esmeralda 118 47181 Wayne 160 05075 Lawrence 201 21001 Adair 93 North Carolina 47101 Lewis 168 California 21137 Lincoln 97 37043 Clay 63 47091 Johnson 169 06063 Plumas 152 21231 Wayne 99 37193 Wilkes 104 47099 Lawrence 172 06033 Lake 199 21057 Cumberland 101 37075 Graham 124 47179 Washington 198 Colorado 21077 Gallatin 108 37023 Burke 176 47177 Warren 203 08025 Crowley 220 21011 Bath 125 37039 Cherokee 189 47095 Lake 216 Georgia 21085 Grayson 126 Ohio Texas 13111 Fannin 82 21089 Greenup 129 39001 Adams 51 48155 Foard 204 13281 Towns 120 21087 Green 132 39131 Pike 72 Utah 13213 Murray 159 21045 Casey 153 39079 Jackson 111 49007 Carbon 84 13143 Haralson 200 21043 Carter 154 39105 Meigs 123 49001 Beaver 114 Illinois 21171 Monroe 163 39015 Brown 127 49015 Emery 186 17069 Hardin 68 21079 Garrard 167 39145 Scioto 136 Vermont Indiana 21201 Robertson 175 39163 Vinton 146 50009 Essex 143 18143 Scott 32 21135 Lewis 178 39053 Gallia 155 50025 Windham 219 18175 Washington 57 21061 Edmonson 179 39009 Athens 173 Virginia 18149 Starke 70 21003 Allen 180 39027 Clinton 190 51027 Buchanan 28 18041 Fayette 81 21019 Boyd 187 39071 Highland 196 51051 Dickenson 29 18155 Switzerland 94 21105 Hickman 191 Oklahoma 51167 Russell 61 18025 Crawford 112 21027 Breckinridge 202 40067 Jefferson 89 51105 Lee 73 18065 Henry 128 21037 Campbell 212 40025 Cimarron 217 51195 Wise 78 18079 Jennings 158 21167 Mercer 214 Pennsylvania 51185 Tazewell 96 18137 Ripley 195 Maine 42079 Luzerne 38 51141 Patrick 166 18029 Dearborn 213 23027 Waldo 135 42021 Cambria 131 51197 Wythe 210 Kansas 23025 Somerset 145 42039 Crawford 188 West Virginia 20207 Woodson 144 23029 Washington 170 Tennessee 54047 McDowell 2 20001 Allen 171 23011 Kennebec 193 47067 Hancock 13 54059 Mingo 7 20205 Wilson 181 Michigan 47087 Jackson 19 54109 Wyoming 16 20153 Rawlins 218 26129 Ogemaw 86 47005 Benton 24 54081 Raleigh 18 Kentucky 26035 Clare 87 47151 Scott 26 54045 Logan 20 21237 Wolfe 1 26135 Oscoda 88 47135 Perry 33 54005 Boone 22 21025 Breathitt 3 26119 Montmorency 91 47071 Hardin 36 54019 Fayette 27 21193 Perry 4 26085 Lake 137 47029 Cocke 41 54065 Morgan 44 21051 Clay 5 26141 Presque Isle 174 47015 Cannon 42 54063 Monroe 47 21013 Bell 6 26001 Alcona 184 47137 Pickett 43 54029 Hancock 49 21131 Leslie 8 26143 Roscommon 192 47013 Campbell 46 54015 Clay 60 21121 Knox 9 26039 Crawford 197 47019 Carter 59 54099 Wayne 62 21071 Floyd 10 26079 Kalkaska 207 47027 Clay 64 54009 Brooke 76 21053 Clinton 11 26031 Cheboygan 215 47057 Grainger 66 54053 Mason 85 21189 Owsley 12 Mississippi 47073 Hawkins 71 54013 Calhoun 90 21235 Whitley 14 28141 Tishomingo 164 47173 Union 74 54067 Nicholas 98 21197 Powell 15 Missouri 47059 Greene 79 54089 Summers 110 21119 Knott 17 29179 Reynolds 55 47025 Claiborne 80 54101 Webster 113 21195 Pike 21 29123 Madison 58 47085 Humphreys 83 54043 Lincoln 121 21153 Magoffin 23 29187 St. Francois 69 47145 Roane 92 54011 Cabell 122 21065 Estill 25 29039 Cedar 107 47133 Overton 95 54091 Taylor 133 21129 Lee 30 29093 Iron 117 47041 DeKalb 102 54055 Mercer 147 21165 Menifee 31 29223 Wayne 119 47143 Rhea 103 54007 Braxton 150 21159 Martin 34 29221 Washington 130 47121 Meigs 105 54095 Tyler 162 21021 Boyle 35 29055 Crawford 148 47129 Morgan 106 54087 Roane 165 21127 Lawrence 39 29085 Hickory 156 47049 Fentress 115 54051 Marshall 182 21203 Rockcastle 40 29013 Bates 177 47111 Macon 116 54003 Berkeley 205 21095 Harlan 45 29181 Ripley 183 47185 White 134 54039 Kanawha 209 21147 McCreary 48 8

Table S2. States with at least one county identified in the top 5% of highest vulnerability ranks by number of vulnerable counties and percentage of all state counties identified as vulnerable. State Vulnerable # Total # Identified Vulnerable (%) Alabama 4 67 6.0 Arizona 1 15 6.7 Arkansas 2 75 2.7 California 2 58 3.5 Colorado 1 64 1.6 Georgia 4 159 2.5 Illinois 1 102 1.0 Indiana 10 92 10.9 Kansas 4 105 3.8 Kentucky 54 120 45.0 Maine 4 16 25.0 Michigan 11 83 13.3 Mississippi 1 82 1.2 Missouri 13 115 11.3 Montana 2 56 3.6 Nevada 2 17 11.8 North Carolina 5 100 5.0 Ohio 11 88 12.5 Oklahoma 2 77 2.6 Pennsylvania 3 67 4.5 Tennessee 41 95 43.2 Texas 1 254 0.4 Utah 3 29 10.3 Vermont 2 14 14.3 Virginia 8 134 6.0 West Virginia 28 55 50.9 Total 220 2139 10.3% Figures Figure S1. County-level indicators investigated for association with acute HCV infection. 9

10

11

12

13

14

15

16

17

Figure S2. Acute HCV infection rate by county. Reported rate of acute HCV infection by county, NNDSS 2012-2013 and model-estimated rate of acute HCV infection by county 18

Figure S3. A. Histogram showing vulnerability scores by number of counties in each score group (eg, 0 to 0.25), B. Sigmoid curve showing vulnerability scores by county rank and identifying the scores bordering the top 5% cut-off, and C. Caterpillar curve of vulnerability scores by mean rank 90% confidence intervals bordering the top 5% rank cut-off (red dash line). 19